by Chris DePuy, 650 Group / October 27, 2026
Akamai launched Akamai Inference and Akamai Agent Sandbox in beta today, adding managed model serving and an isolated execution environment for AI agents to Akamai Cloud. We think the launch is worth attention because Akamai has an interesting advantage in offering inference services: an extensive geographic presence with its many Points of Presence (POPs) — more than 4,350 edge locations, per the announcement — and a very strong track record in security services, a market 650 Group covers in numerous reports. As 650 Group has been tracking in its AI infrastructure research, the companies that can serve models close to users — and protect them — hold a position the hyperscalers do not.
Why does Akamai’s footprint matter for AI inference?
Akamai built its network to move content close to users, and its compute now spans a continuum — GPU- and CPU-dense clusters in the core, thinning out across more than 4,350 edge locations — on one interconnected network fabric. For inference — especially real-time, interactive workloads — placement against the user is the latency lever, and it is the asset a CDN operator brings to AI infrastructure that a region-concentrated cloud does not. The company reports more than $14 billion in multiyear Akamai Cloud contracts as of October 2026.
Which option fits which workload depends on control requirements: teams that need an AI feature live this quarter without operating a serving stack can use the serverless, pay-per-use tier; teams that need control over model versions, deprecation timelines, and tenancy can use dedicated Blackwell Ultra GPU clusters; teams running agents that write their own code need the Sandbox isolation layer regardless.
Akamai Inference: Managed Model Serving on OpenAI-Compatible APIs
Akamai Inference is a managed serving layer for open-weight and custom models, exposed through OpenAI-compatible APIs — an application already written against another provider’s endpoint moves over without rewrites. Two commercial shapes are on offer: serverless pay-per-use, or dedicated deployments. The launch catalog carries 12 open-weight models spanning 8 billion to 2.8 trillion parameters — NVIDIA Nemotron 3 Ultra, OpenAI gpt-oss-120b, and Google Gemma 4 among them — plus embedding models for retrieval-augmented generation. Data grounding runs through Akamai Cloud’s managed databases and object storage, with Akamai operating the orchestration, serving, and authentication layers. We think the OpenAI-compatible API choice is table stakes for enterprise inference in 2026 — the rewrite cost is the adoption gate — and the catalog’s span aims at both feature-level and frontier-class serving. The infrastructure bet is not new: in October 2025, Akamai introduced Akamai Inference Cloud with NVIDIA, pairing RTX PRO 6000 Blackwell Server Edition GPUs and BlueField-3 DPUs with its distributed platform; today’s launch adds the managed service layer on top.
Akamai Agent Sandbox: A Hard Boundary Around Agent-Written Code
AI agents increasingly write and run their own code, and that code needs a boundary. Akamai Agent Sandbox separates the two: the agent itself runs on Akamai Functions, and any code the agent writes executes apart from that runtime. Akamai delivers the component as an open-source part of the Spin framework, and WebAssembly isolation caps the compute, memory, outbound connections, and system capabilities the generated code can touch. The company’s map of the isolation market lines up three converging designs — microVMs with snapshot-and-fork, policy-enforced containers, and WebAssembly-isolated functions — and Akamai leaves the architecture choice to the developer across edge and core. Agent-governance controls, covering what an agent can touch and what can touch the agent, are on the roadmap. We think the sandbox is the sharper strategic product of the two: inference serving has many suppliers, but a security franchise that can enforce boundaries around agent-written code is a rarer combination.
A Broader GPU Fleet: Blackwell Ultra Now, Vera Rubin in 2027
On the fleet side, Akamai added NVIDIA Blackwell Ultra GPU clusters for dedicated use and is readying its data centers for the NVIDIA Vera Rubin platform in 2027. The fleet spans RTX PRO 6000 Blackwell Server Edition GPUs in low-latency serving roles and Blackwell Ultra clusters aimed at teams that train, tune, and host models themselves. Memory is the headline: 288 GB of HBM3e per GPU and 2.3 TB across each eight-GPU Blackwell Ultra node — capacity that holds frontier-class models and suits the memory-hungry workload classes the company calls out: test-time scaling, reasoning chains, and long-context agent work. Dedicated clusters carry no rate limits and no shared-tenancy queueing, attach to the platform’s common data, networking, and security services, and are available through Akamai Sales.
Security Carries the P&L
The security track record is not an anecdote; it is the largest line of Akamai’s income statement. In the second quarter of 2026, security revenue was $604.4 million, up 10% Y/Y — roughly 55% of the company’s $1.1 billion total revenue — per Akamai’s 10-Q filing. Delivery and other cloud applications declined 6% to $395.9 million, while cloud infrastructure services — the line this AI stack runs on — grew 39% Y/Y to $99.3 million (as of Q2 2026). 650 Group’s recent coverage points the same direction: Fortinet’s inline AI security push, Proofpoint’s $6.3 billion acquisition of Varonis, and Cloudflare’s AI workspace partnerships all track the pull between AI workloads and the security layer around them. Akamai stacks serving, execution, and security on one platform rather than leaving enterprises to secure the interaction points between separate providers — inference routing as a network layer is the same movement from the routing side.
The edge-plus-security operator moving up the AI infrastructure stack is a clean, multi-category story: model serving, agent execution, and security sold on one global network, against hyperscalers concentrated in a few dozen regions and inference specialists without a security franchise. Both products are in beta now, with B300 cluster availability running through Akamai Sales. Overall, we think the more telling number in Akamai’s quarter was cloud infrastructure services growing 39% while delivery declined 6%: the platform’s center of gravity is shifting from moving content to running workloads, and today’s launch is the managed-service layer for that shift. 650 Group researches several of the technologies we covered in this blog, including Security and SASE, Cloud and Data Center, and more.
